Clinical Prediction Model: Multisystem Inflammatory Syndrome in Children versus Kawasaki Disease

Abstract

Background:

Multisystem inflammatory syndrome in children (MIS-C) is a rare but serious complication of severe acute respiratory syndrome coronavirus 2 infection. Features of MIS-C overlap with those of Kawasaki disease (KD).

Objective:

The study objective was to develop a prediction model to assist with this diagnostic dilemma.

Methods:

Data from a retrospective cohort of children hospitalized with KD prior to the coronavirus disease 2019 pandemic were compared to a prospective cohort of children hospitalized with MIS-C. A bootstrapped backwards selection process was used to develop a logistic regression model predicting the probability of MIS-C diagnosis. A nomogram was created for application to individual patients.

Results:

Compared to children with incomplete and complete KD (N=602), children with MIS-C (N=105) were older and had longer hospitalizations; more frequent intensive care unit admissions and vasopressor use; lower white blood cell count, lymphocyte count, erythrocyte sedimentation rate, platelet count, sodium, and alanine aminotransferase; and higher hemoglobin and C-reactive protein (CRP) at admission. Left ventricular dysfunction was more frequent in patients with MIS-C, whereas coronary abnormalities were more common in those with KD. The final prediction model included age, sodium, platelet count, alanine aminotransferase, reduction in left ventricular ejection fraction, and CRP. The model exhibited good discrimination with AUC 0.96 (95% CI: [0.94-0.98]) and was well calibrated (optimism-corrected intercept of −0.020 and slope of 0.99).

Conclusions:

A diagnostic prediction model utilizing admission information provides excellent discrimination between MIS-C and KD. This model may be useful for diagnosis of MIS-C but requires external validation.

Introduction

Children with coronavirus disease 2019 (COVID-19) generally have mild symptoms.¹ However, some develop a hyperinflammatory condition, termed multisystem inflammatory syndrome in children (MIS-C), within weeks of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infection.² Diagnosing MIS-C is challenging as features overlap with other hyperinflammatory conditions, including Kawasaki disease (KD).^3,4 Yet, it is imperative to distinguish MIS-C from KD because complications, long-term sequelae, and treatments differ between them.

Many studies have compared the clinical features of MIS-C and KD. The majority included data from a single-center and/or had small sample sizes; few included over 100 patients within each group.^5-24 Generally, these reports suggested that patients with MIS-C tended to be older, required intensive care more often, and had more frequent gastrointestinal involvement compared to patients with KD. Prior studies also associated MIS-C with thrombocytopenia, lymphopenia, myocarditis, and decreased ventricular function, while KD was associated with leukocytosis and coronary artery involvement. Further, inflammatory markers, including C-reactive protein (CRP), tended to be higher in children with MIS-C.^{7,11,13,17-20,22-24} Despite these differences, distinguishing MIS-C from KD remains clinically challenging.

Limited prediction models exist distinguishing MIS-C from KD. Godfred-Cato et al. studied 233 patients with MIS-C and 101 with KD and developed a risk score that included thrombocytopenia, abdominal pain, headache, pericardial effusion, elevated CRP, rash, and mucocutaneous involvement.²¹ This model was internally validated and had an AUC of 0.96. Lam et al. developed a machine-learning algorithm using age, KD criteria, and 17 laboratory measurements that had similar performance both in internal development and external validation.²⁵ Another model from Clark et al. used hypotension/fluid resuscitation, abdominal pain, rash, and sodium to predict a diagnosis of MIS-C, although not specifically compared to KD.²⁶ While useful in many settings, these models are limited by the use of subjective symptoms^21,26 and complex dependence on numerous labs.²⁵ Further, their generalizability is not known. Therefore, we developed and internally validated a prediction model to distinguish between children with MIS-C and KD seen at a large referral center in the mid-South. We sought to develop a model that used objective information to predict these diagnoses in order to improve ease of use.

Methods

Study design and population

We conducted a retrospective chart review among children hospitalized at Monroe Carell Jr. Children’s Hospital at Vanderbilt (MCJCH) between May 29, 2000 and December 3, 2019 with a diagnosis of KD. This time period was chosen to avoid overlap with MIS-C cases. We identified children with KD based on International Classification of Diseases, Ninth Edition, Clinical Modification code 446.1 (acute febrile mucocutaneous lymph node syndrome), Internal Classification of Diseases, Tenth Edition (ICD-10), Clinical Modification Code M30.3 (mucocutaneous lymph node syndrome), or electronic medical records that included ≥5 mentions of KD based on a standardized Vanderbilt research derivative.²⁷ Children were included if they: (1) met the American Heart Association diagnostic criteria for complete or incomplete KD (Supplementary Methods);²⁸ (2) received at least one dose of intravenous immunoglobulin (IVIG); (3) were less than 18 years old; and (4) were discharged with a primary diagnosis of KD. Children were excluded if they had an alternative diagnosis at follow-up.

Prospective active surveillance was conducted for all patients with MIS-C seen at MCJCH between July 11, 2020 and September 18, 2022. Patients were included if they: (1) were hospitalized at MCJCH; (2) were discharged with a primary diagnosis of MIS-C; and (3) met the 2023 Centers for Disease Control and Prevention (CDC) case definition for MIS-C (Supplementary Methods).²⁹ Data were collected contemporaneously during their hospitalization.

Both KD and MIS-C protocols were approved by the Vanderbilt Institutional Review Board. We used the methodological outline of the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) Guidelines in developing the predictive model.³⁰

Data collection

Data were abstracted from electronic medical records and entered into a REDCap database.^31,32 Demographic variables included age, sex, and self-reported race and ethnicity. KD diagnostic criteria²⁸ reported by family or providers included fever, conjunctivitis, changes in the lips and oral cavity (mucositis), cervical lymphadenopathy, extremity changes, and rash. Laboratory data included white blood cell (WBC) count, hemoglobin, CRP, erythrocyte sedimentation rate (ESR), absolute lymphocyte count (ALC), absolute neutrophil count (ANC), albumin, platelet count, sodium, aspartate aminotransferase (AST), and alanine aminotransferase (ALT) collected within the first 24 hours of admission.

Echocardiographic findings at presentation were also abstracted, including any abnormality, coronary artery ectasia and/or aneurysm, left ventricular ejection fraction (LVEF), mitral regurgitation, and pericardial effusion. We defined reduced LVEF as less than 55% measured using the bullet method as per the recommendation by the American Society of Echocardiography.³³ We assessed the degree of mitral regurgitation using a five-category scale (none, trivial, mild, moderate, severe). Severity measures included length of hospital stay in days, intensive care unit (ICU) admission, and vasopressor use.

Descriptive analysis

We summarized categorical variables as frequency (percentage) and continuous variables as mean (standard deviation [SD]). Comparisons were made between complete KD and MIS-C patients and incomplete KD and MIS-C patients using Pearson’s chi-squared for categorical variables and two-sample t-tests with unequal variances for continuous variables. A significance level of 0.05 was used for all analyses (two-tailed, where appropriate).

Outcomes and candidate predictors

The outcome for our prediction model was a final diagnosis of MIS-C or KD. We chose candidate predictors a priori based on our clinical experience and previous studies.^5-24 The following twelve candidate predictors were considered for inclusion in the model: age, vasopressor use, sodium, albumin, ALC, ANC, CRP, AST, ALT, platelet count, LVEF (normal or reduced), and coronary involvement. Sample size calculations were performed using sample size criteria from Riley et al.³⁴

Model development and validation

Missing values were first imputed by predictive mean matching using the aregImpute function (Hmisc package in R) using 10 imputations. We selected 20% as the threshold for the allowable degree of missingness. The variable albumin had a high frequency of missingness (28%). Data reduction techniques including redundancy analysis and hierarchical clustering using Hoeffding’s D statistic were performed to determine if collinearity or redundancy existed between the candidate predictors.³⁵ Continuous predictors were plotted to visually assess linearity.

We used a non-parametric bootstrapped backwards selection process to determine the most important predictors for the final logistic regression model.³⁶ In this process, 500 bootstrap samples were drawn from the original data, and backwards selection was performed in each. The choice of predictors in the final model was determined by the frequency with which they were selected across the 500 bootstrapped samples; those that were selected in at least 80% of the bootstrapped models were included in the final model. A heuristic shrinkage factor was applied in the final model to account for overfitting and improve predictions in new patients by shrinking regression coefficients.^35,37

The predictive accuracy of the model was evaluated using a calibration curve, Brier score, and area under the receiver operating characteristic curve (AUC).^35,38 To correct for model optimism, the model was internally validated using a bootstrap procedure.³⁵ A nomogram was created to visually represent the calculation of individual predicted values and risks.

We conducted sensitivity analyses comparing the performance of the model with and without albumin due to its frequency of missingness. The analysis was repeated across the 10 imputed datasets. Finally, we used the model to make outcome predictions for children with KD hospitalized at MCJCH during the same date interval as patients with MIS-C in our cohort. To identify cases, we queried the Pediatric Health Information System^® database for all patients admitted at our site between July 11, 2020, and September 18, 2022, with a principal ICD-10 diagnosis code of M30.3. We manually reviewed electronic medical records to exclude cases that did not meet KD criteria and collect data on variables included in the final model. All analyses were performed using R software version 4.1.1.

Results

Demographics

During the study period and prior to the SARS-CoV-2 pandemic, 602 children were hospitalized at MCJCH with KD, of which 218 (36.2%) had incomplete KD. During the SARS-CoV-2 pandemic study period, 105 individuals were seen at MCJCH with MIS-C. The mean age was younger for children with complete KD (3.6 years [SD=2.7], p<0.001) and incomplete KD (3.7 years [SD=3.1], p<0.001) compared to MIS-C (9.5 years [SD=4.3]; Table 1). No differences in sex and race or reported Hispanic ethnicity between groups were noted. Missing variables for individual children are summarized in Supplementary Table 1.

Table 1.

Comparison of the clinical characteristics of children with MIS-C and those with complete or incomplete KD seen at Monroe Carell Jr. Children's Hospital at Vanderbilt

Characteristic	MIS-C, n=105	Complete KD, n=384	p - value*	Incomplete KD, n=218	p - value*
Demographics
Age (years)—mean (SD)	9.5 (4.3)	3.6 (2.7)	<0.001	3.7 (3.1)	<0.001
Age group (years)—n (%)			<0.001		<0.001
<5	15 (14.3)	289 (75.3)		162 (74.3)
5–10	57 (54.3)	88 (22.9)		47 (21.6)
11–17	32 (30.5)	7 (1.8)		9 (4.1)
18–20	1 (1.0)	0 (0.0)		0 (0.0)
Male—n (%)	67 (63.8)	243 (63.3)	>0.99	144 (66.1)	0.79
Race and ethnicity —n (%)			0.38		0.31
White non-Hispanic	48/98 (49.0)	217/379 (57.3)		123/215 (57.2)
Black non-Hispanic	26/98 (26.5)	96/379 (25.3)		58/215 (27.0)
Hispanic	14/98 (14.3)	39/379 (10.3)		19/215 (8.8)
Other non-Hispanic	10/98 (10.2)	27/379 (7.1)		15/215 (7.0)
Kawasaki Diagnostic Criteria**
Fever—n (%)	105 (100.0)	384 (100.0)		218 (100.0)
Rash—n (%)	69 (65.7)	379 (98.7)	<0.001	156 (71.6)	0.35
Mucositis—n (%)	58 (55.2)	374 (97.4)	<0.001	136 (62.4)	0.27
Conjunctivitis—n (%)	77 (73.3)	376 (97.9)	<0.001	173 (79.4)	0.28
Cervical lymphadenopathy—n (%)	43/103 (41.7)	148 (38.5)	0.63	33/218 (15.1)	<0.001
Extremity changes—n (%)	34 (32.4)	333 (86.7)	<0.001	77 (35.3)	0.69
Number of Kawasaki Diagnostic Criteria**
Fever only—n (%)	8 (7.6)	-	-	4 (1.8)	-
Fever + 1 criterion—n (%)	19 (18.1)	-	-	10 (4.6)	-
Fever + 2 criteria—n (%)	19 (18.1)	-	-	47 (21.6)	-
Fever + 3 criteria—n (%)	23 (21.9)	-	-	157 (72.0)	-
Fever + 4 criteria—n (%)	25 (23.8)	310 (80.7)	-	-	-
Fever + 5 criteria—n (%)	11 (10.5)	74 (19.3)	-	-	-
Clinical Course
Days in hospital—mean (SD)	5.2 (2.9)	3.4 (2.7)	<0.001	4.0 (3.2)	0.001
ICU admission—n (%)	47 (44.8)	6/382 (1.6)	<0.001	16 (7.3)	<0.001
Vasopressor use—n (%)	20 (19.0)	2 (0.5)	<0.001	7 (3.2)	<0.001
Laboratory Values
WBC count (×103/μL)—mean (SD)	10.5 (4.9)	14.4 (5.6)	<0.001	14.2 (5.9)	<0.001
Hemoglobin (g/dL)—mean (SD)	11.6 (1.5)	10.9 (1.3)	<0.001	10.6 (1.3)	<0.001
CRP (mg/L)—mean (SD)	162.8 (91.2)	111.9 (79.0)	<0.001	109.1 (85.9)	<0.001
ESR (mm/hour)—mean (SD)	44.9 (23.4)	62.6 (30.7)	<0.001	65.0 (31.6)	<0.001
ALC (×103/μL)—mean (SD)	1.1 (0.9)	3.0 (2.0)	<0.001	3.4 (2.4)	<0.001
ANC (×103/μL)—mean (SD)	8.5 (4.2)	9.8 (4.9)	0.008	9.4 (5.2)	0.11
Albumin (g/dL)—mean (SD)	3.4 (0.5)	3.4 (0.5)	0.70	3.3 (0.5)	0.024
Platelet count (×103/μL)—mean (SD)	182.7 (91.9)	392.3 (166.8)	<0.001	399.0 (194.8)	<0.001
Sodium (mmol/L)—mean (SD)	132.6 (3.4)	135.8 (3.0)	<0.001	136.2 (2.7)	<0.001
AST (U/L)—mean (SD)	58.6 (97.2)	55.3 (75.3)	0.76	67.5 (108.9)	0.48
ALT (U/L)—mean (SD)	46.5 (79.4)	67.2 (75.4)	0.02	67.8 (92.7)	0.04
Echocardiogram Findings
Abnormal echocardiogram—n (%)	79/103 (76.7)	166 (43.2)	<0.001	117 (53.7)	<0.001
Coronary artery ectasia—n (%)	6/103 (5.8)	63 (16.4)	0.01	51 (23.4)	<0.001
LVEF <55%—n (%)	52/103 (50.5)	12 (3.1)	<0.001	17/218 (7.8)	<0.001
Mitral regurgitation—n (%)			<0.001		<0.001
None	48/102 (47.1)	297 (77.3)		161/218 (73.9)
Trivial	38/102 (37.3)	66 (17.2)		37/218 (17.0)
Mild	13/102 (12.7)	17 (4.4)		14/218 (6.4)
Moderate	3/102 (2.9)	4 (1.0)		6/218 (2.8)
Severe	0/102 (0.0)	0 (0.0)		0/218 (0.0)
Pericardial effusion—n (%)	10/103 (9.7)	64 (16.7)	0.11	39/218 (17.9)	0.08

^*p-values were calculated using Pearson’s χ² test for categorical variables and the two-sample t-test with unequal variances for continuous variables. Table p-values reflect comparisons between MIS-C to complete KD and MIS-C to incomplete KD.

^**Kawasaki disease criteria refers to the American Heart Association diagnostic criteria.²⁸ Abbreviations: ALC, absolute lymphocyte count; ALT, alanine aminotransferase; ANC, absolute neutrophil count; AST, aspartate aminotransferase; CRP, C-reactive protein; ESR, erythrocyte sedimentation rate; ICU, intensive care unit; KD, Kawasaki disease; LVEF, left ventricular ejection fraction; MIS-C, multisystem inflammatory disorder in children; SD, standard deviation; WBC, white blood cell.

Clinical characteristics and outcomes

Comparisons of KD diagnostic criteria between patients with MIS-C, incomplete KD, and KD are in Table 1.

Patients in the MIS-C cohort had lower WBC count, ALC, platelet count, ESR, sodium, and ALT and higher hemoglobin and CRP compared to both KD groups (all p-values<0.05). Children with MIS-C had higher albumin (p=0.03) than those with incomplete KD.

Children with MIS-C had longer hospital stays, more frequent ICU admission, and more frequent vasopressor use than children with both complete and incomplete KD (all p-values<0.002).

Cardiac variables

Echocardiogram abnormalities were more common in children with MIS-C (76.7%) than in those with complete (43.2%, p<0.001) and incomplete KD (53.7%, p<0.001) (Table 1). Reduced LVEF was more common in children with MIS-C (50.5%) than in those with complete (3.1%, p<0.001) and incomplete KD (7.8%, p<0.001). Mitral regurgitation was more common in children with MIS-C; more than half (52.9%) had at least trivial regurgitation (p<0.001). Coronary artery ectasia was more common in children with complete and incomplete KD (16.4% and 23.4%, respectively) than those with MIS-C (5.8%; p=0.01 and p<0.001, respectively).

Prediction model

The sensitivity analysis did not show marked changes with or without albumin; albumin was therefore removed from further analyses due to missingness. The assessment of linearity suggested linearity to be a reasonable approximation for the purposes of model building. The bootstrapped backwards selection process indicated the most important predictors for discriminating MIS-C from KD were age, platelet count, sodium, ALT, CRP, and LVEF (Supplementary Table 2). The full multivariable prediction model is presented in Table 2. The AUC was 0.96 (95% CI: [0.94-0.98]) (Figure 1), showing the model was very effective at distinguishing MIS-C cases. Supplementary Figure 1 indicates observed and predicted outcomes agreed closely. Model metrics indicate a close-to-ideal bias-corrected intercept and slope, and high model accuracy. The c-statistic indicated good internal validation. These results were consistent in sensitivity analyses comparing 10 imputed datasets.

Table 2.

Final multivariable logistic regression model predicting the risk of MIS-C among 602 children with complete or incomplete KD and 105 children with MIS-C

Predictor	β coefficient	Odds Ratio	95% CI	p-value
Intercept	30.802
Age (years)	0.309	1.363	(1.246, 1.489)	<0.001
Sodium (mmol/L)	−0.242	0.785	(0.706, 0.873)	<0.001
Platelet count (×103/μL)	−0.009	0.991	(0.988, 0.994)	<0.001
ALT (U/L)*	−0.005	0.995	(0.989, 1.001)	0.104
LVEF <55%	1.523	4.586	(2.217, 9.485)	<0.001
CRP (mg/L)	0.004	1.004	(1.000**, 1.007)	0.046

Abbreviations: ALT, alanine aminotransferase; CI, confidence interval; CRP, C-reactive protein; KD, Kawasaki disease; LVEF, left ventricular ejection fraction; MIS-C, multi-system inflammatory syndrome in children.

^*Though ALT does not reach clinical significance in the model, it was included per the bootstrapped backwards selection technique in 93.6% of bootstrapped iterations (Supplementary Table 2). This process identifies the most important predictors for MIS-C, thus all variables that are selected in over 80% of the 500 bootstraps are included in the final model. Removing these variables would not improve model performance, as they are important for predicting MIS-C risk.

^**The lower limit of the 95% CI for CRP is 1.00007.

Figure 1. Receiver operating characteristic (ROC) curve for prediction of MIS-C.

The ROC curve measures the discrimination of the model. The AUC is 0.96 (95% CI: [0.94-0.98]), indicating the model differentiates well between those with and without MIS-C.

After applying the heuristic shrinkage factor, the model can be applied to individuals using the nomogram or model formula to determine the risk of MIS-C (Figure 2). An example to utilize the nomogram is outlined here. A new MIS-C patient is identified, with age 5.1 years, sodium level 132 mmol/L, platelet count 175×10³/μL, ALT 38 U/L, CRP 161 mg/L, and LVEF <55%. Using the coefficients from the model, we would calculate the risk of MIS-C as: 1/(1 + exp[−(30.802 + 0.309*5.1 – 0.242*132 – 0.009*175 – 0.005*38 + 0.004*161 + 1.523*1)*0.982] = 0.694 (where 0.982 is the heuristic shrinkage factor). Using the nomogram, we will calculate the points for each predictor, then calculate total points. Age 5.1 = 13 points; sodium 132 = 23 points; platelets 175 = 85 points; LVEF reduced = 13 points; ALT 38 = 31 points; CRP 161 = 4 points. Total points: 13+23+85+13+31+4 = 169 points. Drawing from total points to risk of MIS-C, we see the risk is roughly 0.69. This example is also depicted in Supplementary Figure 2. Additional information for using the nomogram can be found in Supplementary Methods.

Figure 2. A nomogram for predicting an individual’s risk of MIS-C.

The nomogram can be applied to individual cases to determine the risk of MIS-C. The following is the formula for risk of MIS-C: 1/(1 + exp("Xβ*0.982")), "Xβ" = 30.802 + 0.309*(age, years) − 0.242*(sodium, mmol/L) − 0.009*(platelet count, ×10³/μL) + 1.523*(LVEF <55%) − 0.005*(ALT, U/L) + 0.004*(CRP, mg/L).

The distribution of predictions made by the model using our cohort can be found in Supplementary Figure 3. Using the model to make outcome predictions for children with KD hospitalized at MCJCH between July 11, 2020, and September 18, 2022, we found that the predicted risk of MIS-C was less than 5% for the majority of the 73 cases identified (n=58; median risk, 1.5%; interquartile range, 0.4–4.1%), with a maximum risk among this group being 49.1%. Therefore,, the model consistently differentiated KD from MIS-C among an independent cohort of children with clinically confirmed KD.

The pmsampsize package in R was used to calculate the minimum sample size required to develop a multivariable prediction model with a 15% outcome rate and between six and 13 predictors. Using the criteria from Riley et al.,³⁴ we estimated between 196 to 356 patients with 30 to 54 outcomes would be required, at minimum, to precisely estimate the average risk of MIS-C. Therefore, our sample size was sufficient to develop this prediction model.

Discussion

We developed and internally validated a prediction model to aid in differentiating KD and MIS-C using demographic information, routinely collected labs within the first 24 hours of admission, and echocardiographic findings at admission. Our model demonstrated great performance and discrimination of MIS-C, enabling an individual’s likelihood of MIS-C to be calculated using the risk formula or nomogram.

The final prediction model identified six commonly available measures at the time of diagnosis to accurately differentiate children with MIS-C from those with KD. In the validation cohort, our model consistently differentiated KD from MIS-C in a population of patients with KD who were hospitalized during the pandemic, suggesting its applicability during a period in which KD and MIS-C overlap.

Our AUC of 0.96 is similar to the AUC recently reported by Godfred-Cato et al.²¹ Their model also included platelets but featured clinical symptoms, including abdominal pain and headache. These were not included in our cohort as they were not systematically captured and can be subjective. Instead, our model includes objective data to guide clinical decision making. Their model also included rash and mucocutaneous lesions, which are diagnostic criteria for KD. Given that these can occur in both clinical entities, this can pose challenges to clinicians attempting to use this model in real-world settings. Interestingly, they found pericardial effusion to be more common in MIS-C and included this in their model, a finding we did not reproduce. Combination of these models may increase accuracy and improve generalizability. Our AUC is also similar to that reported by Lam et al. using a machine learning model.²⁵ However, their model used more predictors such as age, KD diagnostic criteria, and 17 laboratory values, several of which are not routinely collected at our institution. This suggests a lack of generalizability of their model. The model proposed by Clark et al. to distinguish MIS-C from non-MIS-C diagnoses also includes subjective clinical features not collected in our KD cohort and does not specifically compare MIS-C and KD.²⁶ Their model includes rash as a predictor; as a feature of both MIS-C and KD, the model may be limited in distinguishing the two diagnoses. Finally, we used the 2023 CDC case definition²⁹ to identify patients with MIS-C, whereas other models relied on earlier definitions.

Individuals with MIS-C in our cohort were older than children with KD. This is consistent with the literature and may guide clinicians when determining a diagnosis, especially when the majority of children with KD are less than five years.^{5,7-11,13-15,15,17-22} Further understanding of why these age differences exist is needed. The distribution of race and Hispanic ethnicity in our study did not differ between groups. Though previous studies have shown similar results,^5,8,9,11,14 one study reported that KD was more common than MIS-C in Hispanic, non-Hispanic Black, and non-Hispanic Asian children,²¹ while another showed that KD predominantly affected non-Hispanic White children.¹⁷ Further studies are needed to consolidate these contradictory findings.

Most patients with MIS-C did not meet the complete KD diagnostic criteria, with just 10.5% exhibiting all five and only a quarter meeting four. This is consistent with other studies, suggesting that the presence of all five criteria should prompt consideration of KD over MIS-C.^5,15,18-21 Besides cervical lymphadenopathy, there were no other differences in presence of diagnostic criteria between incomplete KD and MIS-C, highlighting the need for a diagnostic model.

Consistent with previous studies, lymphocytopenia was more common in children with MIS-C.^{7-13,15,17-19,21,23} Children with complete or incomplete KD were more likely than those with MIS-C to present with supplemental laboratory findings associated with KD, including higher platelet count, lower hemoglobin, and higher ALT. Similarly, albumin was lower in children with incomplete KD compared to those with MIS-C. We also found that children with MIS-C presented with lower sodium. Several studies also found lower sodium,^{7,11,13,20,23} while others did not find a significant difference.^8-10,17 The variability in laboratory findings poses diagnostic challenges; therefore, a multivariable predictive model is useful.

Our MIS-C cohort had a more severe clinical course compared to the KD cohort, defined by longer hospital stays, greater vasopressor use, and more ICU admissions. Children with MIS-C also had a more pronounced acute inflammatory response than those with KD, as evidenced by a higher CRP in our study and others.^{7,11,13,17-24}

Cardiac involvement is prevalent in both conditions but was more common in MIS-C compared to KD, with an abnormal echocardiogram in three-quarters compared to nearly half, respectively. These frequencies are consistent with a prior study.¹⁷ The difference was driven by decreased LVEF, which occurred in over half of children with MIS-C but was rare in KD. Children with KD had significantly more coronary involvement than children with MIS-C. The low percentage of coronary involvement in our MIS-C cohort was similar to a prior study from the United States,¹⁴ but not as high as some groups reporting 20-40%.^17,39,40 Mitral regurgitation was more common in MIS-C, and over half of children had at least trivial regurgitation. Small studies have generally found increased valvular insufficiency in MIS-C,^9,18 but studies looking specifically at mitral valve involvement have not found a significant difference.^7,14 Of note, our study did not find that pericardial effusion was more common in MIS-C, which has been reported in some studies.^9,17,20 The differences between our cohorts may be explained by the fact that MIS-C cardiovascular involvement is more typically characterized by systolic dysfunction, a known complication of systemic inflammation. While KD can sometimes impact systolic performance, it has a predilection for coronary artery involvement. Therefore, decreased LVEF is more suggestive of MIS-C, while coronary artery involvement suggests a KD diagnosis.

Our study has several limitations. First, data were collected at a single center, which may affect generalizability. This may be partially mitigated by our large historical KD cohort and relatively large MIS-C cohort. Second, the study is limited by retrospective data collection and reliance on medical chart abstraction. Third, our model has not been externally validated, which will be pursued in future studies. The generalizability of our model is not known, and external validation is necessary prior to implementation in patient care. Fourth, it only distinguishes KD from MIS-C and does not include other conditions that can present similarly. We do not know the performance of the model in settings in which other febrile illnesses cannot be excluded. Fifth, patients may have been misdiagnosed with KD or MIS-C after the onset of the pandemic due to the overlap of features. To help mitigate this, we selected a KD cohort from prior to the pandemic. Finally, we were unable to compare certain variables, such as B-type natriuretic protein and troponin, between groups as these variables were not routinely available for children with KD.

In conclusion, children who meet criteria for complete KD, especially those that have all five KD diagnostic criteria, have a low probability of having MIS-C. Our prediction model has the potential to serve as a practical tool that can be quickly utilized to distinguish KD from MIS-C when the diagnosis is unclear. The model calls for six predictors that are commonly collected, which can be put into the risk formula or nomogram to provide a reliable risk prediction for hospitalized children. However, external validation is needed prior to implementing this model in patient care. Future studies will investigate application of the model to patients hospitalized at other institutions.